func (n *node) ProposeConfChange(ctx context.Context, cc pb.ConfChange) error {
data, err := cc.Marshal()
if err != nil {
return err
}
return n.Step(ctx, pb.Message{Type: pb.MsgProp, Entries: []pb.Entry{{Type: pb.EntryConfChange, Data: data}}})
}
func (c *RemoveCommand) Data2() ([]byte, error) {
req2 := raftpb.ConfChange{
ID: 0,
Type: raftpb.ConfChangeRemoveNode,
NodeID: c.id,
}
return req2.Marshal()
}
// TestMultiNodeStart ensures that a node can be started correctly. The node should
// start with correct configuration change entries, and can accept and commit
// proposals.
func TestMultiNodeStart(t *testing.T) {
ctx, cancel := context.WithCancel(context.Background())
defer cancel()
cc := raftpb.ConfChange{Type: raftpb.ConfChangeAddNode, NodeID: 1}
ccdata, err := cc.Marshal()
if err != nil {
t.Fatalf("unexpected marshal error: %v", err)
}
wants := []Ready{
{
SoftState: &SoftState{Lead: 1, RaftState: StateLeader},
HardState: raftpb.HardState{Term: 2, Commit: 2, Vote: 1},
Entries: []raftpb.Entry{
{Type: raftpb.EntryConfChange, Term: 1, Index: 1, Data: ccdata},
{Term: 2, Index: 2},
},
CommittedEntries: []raftpb.Entry{
{Type: raftpb.EntryConfChange, Term: 1, Index: 1, Data: ccdata},
{Term: 2, Index: 2},
},
},
{
HardState: raftpb.HardState{Term: 2, Commit: 3, Vote: 1},
Entries: []raftpb.Entry{{Term: 2, Index: 3, Data: []byte("foo")}},
CommittedEntries: []raftpb.Entry{{Term: 2, Index: 3, Data: []byte("foo")}},
},
}
mn := StartMultiNode(1, 10, 1)
storage := NewMemoryStorage()
mn.CreateGroup(1, []Peer{{ID: 1}}, storage)
mn.Campaign(ctx, 1)
gs := <-mn.Ready()
g := gs[1]
if !reflect.DeepEqual(g, wants[0]) {
t.Fatalf("#%d: g = %+v,\n w %+v", 1, g, wants[0])
} else {
storage.Append(g.Entries)
mn.Advance(gs)
}
mn.Propose(ctx, 1, []byte("foo"))
if gs2 := <-mn.Ready(); !reflect.DeepEqual(gs2[1], wants[1]) {
t.Errorf("#%d: g = %+v,\n w %+v", 2, gs2[1], wants[1])
} else {
storage.Append(gs2[1].Entries)
mn.Advance(gs2)
}
select {
case rd := <-mn.Ready():
t.Errorf("unexpected Ready: %+v", rd)
case <-time.After(time.Millisecond):
}
}
func (mn *multiNode) ProposeConfChange(ctx context.Context, group uint64, cc pb.ConfChange) error {
data, err := cc.Marshal()
if err != nil {
return err
}
return mn.Step(ctx, group,
pb.Message{
Type: pb.MsgProp,
Entries: []pb.Entry{
{Type: pb.EntryConfChange, Data: data},
},
})
}
// TestMultiNodeProposeConfig ensures that multiNode.ProposeConfChange
// sends the given configuration proposal to the underlying raft.
func TestMultiNodeProposeConfig(t *testing.T) {
mn := newMultiNode(1, 10, 1)
go mn.run()
s := NewMemoryStorage()
mn.CreateGroup(1, []Peer{{ID: 1}}, s)
mn.Campaign(context.TODO(), 1)
proposed := false
var lastIndex uint64
var ccdata []byte
for {
rds := <-mn.Ready()
rd := rds[1]
s.Append(rd.Entries)
// change the step function to appendStep until this raft becomes leader
if !proposed && rd.SoftState.Lead == mn.id {
cc := raftpb.ConfChange{Type: raftpb.ConfChangeAddNode, NodeID: 1}
var err error
ccdata, err = cc.Marshal()
if err != nil {
t.Fatal(err)
}
mn.ProposeConfChange(context.TODO(), 1, cc)
proposed = true
}
mn.Advance(rds)
var err error
lastIndex, err = s.LastIndex()
if err != nil {
t.Fatal(err)
}
if lastIndex >= 3 {
break
}
}
mn.Stop()
entries, err := s.Entries(lastIndex, lastIndex+1, noLimit)
if err != nil {
t.Fatal(err)
}
if len(entries) != 1 {
t.Fatalf("len(entries) = %d, want %d", len(entries), 1)
}
if entries[0].Type != raftpb.EntryConfChange {
t.Fatalf("type = %v, want %v", entries[0].Type, raftpb.EntryConfChange)
}
if !bytes.Equal(entries[0].Data, ccdata) {
t.Errorf("data = %v, want %v", entries[0].Data, ccdata)
}
}
// applyConfChange applies a ConfChange to the server. It is only
// invoked with a ConfChange that has already passed through Raft
func (s *EtcdServer) applyConfChange(cc raftpb.ConfChange, confState *raftpb.ConfState) (bool, error) {
if err := s.Cluster.ValidateConfigurationChange(cc); err != nil {
cc.NodeID = raft.None
s.r.ApplyConfChange(cc)
return false, err
}
*confState = *s.r.ApplyConfChange(cc)
switch cc.Type {
case raftpb.ConfChangeAddNode:
m := new(Member)
if err := json.Unmarshal(cc.Context, m); err != nil {
log.Panicf("unmarshal member should never fail: %v", err)
}
if cc.NodeID != uint64(m.ID) {
log.Panicf("nodeID should always be equal to member ID")
}
s.Cluster.AddMember(m)
if m.ID == s.id {
log.Printf("etcdserver: added local member %s %v to cluster %s", m.ID, m.PeerURLs, s.Cluster.ID())
} else {
s.r.transport.AddPeer(m.ID, m.PeerURLs)
log.Printf("etcdserver: added member %s %v to cluster %s", m.ID, m.PeerURLs, s.Cluster.ID())
}
case raftpb.ConfChangeRemoveNode:
id := types.ID(cc.NodeID)
s.Cluster.RemoveMember(id)
if id == s.id {
return true, nil
} else {
s.r.transport.RemovePeer(id)
log.Printf("etcdserver: removed member %s from cluster %s", id, s.Cluster.ID())
}
case raftpb.ConfChangeUpdateNode:
m := new(Member)
if err := json.Unmarshal(cc.Context, m); err != nil {
log.Panicf("unmarshal member should never fail: %v", err)
}
if cc.NodeID != uint64(m.ID) {
log.Panicf("nodeID should always be equal to member ID")
}
s.Cluster.UpdateRaftAttributes(m.ID, m.RaftAttributes)
if m.ID == s.id {
log.Printf("etcdserver: update local member %s %v in cluster %s", m.ID, m.PeerURLs, s.Cluster.ID())
} else {
s.r.transport.UpdatePeer(m.ID, m.PeerURLs)
log.Printf("etcdserver: update member %s %v in cluster %s", m.ID, m.PeerURLs, s.Cluster.ID())
}
}
return false, nil
}
// StartNode returns a new Node given a unique raft id, a list of raft peers, and
// the election and heartbeat timeouts in units of ticks.
// It appends a ConfChangeAddNode entry for each given peer to the initial log.
func StartNode(id uint64, peers []Peer, election, heartbeat int, storage Storage) Node {
n := newNode()
r := newRaft(id, nil, election, heartbeat, storage, 0)
// become the follower at term 1 and apply initial configuration
// entires of term 1
r.becomeFollower(1, None)
for _, peer := range peers {
cc := pb.ConfChange{Type: pb.ConfChangeAddNode, NodeID: peer.ID, Context: peer.Context}
d, err := cc.Marshal()
if err != nil {
panic("unexpected marshal error")
}
e := pb.Entry{Type: pb.EntryConfChange, Term: 1, Index: r.raftLog.lastIndex() + 1, Data: d}
r.raftLog.append(e)
}
// Mark these initial entries as committed.
// TODO(bdarnell): These entries are still unstable; do we need to preserve
// the invariant that committed < unstable?
r.raftLog.committed = r.raftLog.lastIndex()
r.Commit = r.raftLog.committed
// Now apply them, mainly so that the application can call Campaign
// immediately after StartNode in tests. Note that these nodes will
// be added to raft twice: here and when the application's Ready
// loop calls ApplyConfChange. The calls to addNode must come after
// all calls to raftLog.append so progress.next is set after these
// bootstrapping entries (it is an error if we try to append these
// entries since they have already been committed).
// We do not set raftLog.applied so the application will be able
// to observe all conf changes via Ready.CommittedEntries.
for _, peer := range peers {
r.addNode(peer.ID)
}
go n.run(r)
return &n
}
// configure sends a configuration change through consensus and
// then waits for it to be applied to the server. It
// will block until the change is performed or there is an error.
func (s *EtcdServer) configure(ctx context.Context, cc raftpb.ConfChange) error {
cc.ID = s.reqIDGen.Next()
ch := s.w.Register(cc.ID)
if err := s.r.ProposeConfChange(ctx, cc); err != nil {
s.w.Trigger(cc.ID, nil)
return err
}
select {
case x := <-ch:
if err, ok := x.(error); ok {
return err
}
if x != nil {
log.Panicf("return type should always be error")
}
return nil
case <-ctx.Done():
s.w.Trigger(cc.ID, nil) // GC wait
return parseCtxErr(ctx.Err())
case <-s.done:
return ErrStopped
}
}
func (mn *multiNode) run() {
groups := map[uint64]*groupState{}
rds := map[uint64]Ready{}
var advancec chan map[uint64]Ready
for {
// Only select readyc if we have something to report and we are not
// currently waiting for an advance.
readyc := mn.readyc
if len(rds) == 0 || advancec != nil {
readyc = nil
}
// group points to the group that was touched on this iteration (if any)
var group *groupState
select {
case gc := <-mn.groupc:
// TODO(bdarnell): pass applied through gc and into newRaft. Or get rid of it?
r := newRaft(mn.id, nil, mn.election, mn.heartbeat, gc.storage, 0)
group = &groupState{
id: gc.id,
raft: r,
}
groups[gc.id] = group
lastIndex, err := gc.storage.LastIndex()
if err != nil {
panic(err) // TODO(bdarnell)
}
// If the log is empty, this is a new group (like StartNode); otherwise it's
// restoring an existing group (like RestartNode).
// TODO(bdarnell): rethink group initialization and whether the application needs
// to be able to tell us when it expects the group to exist.
if lastIndex == 0 {
r.becomeFollower(1, None)
ents := make([]pb.Entry, len(gc.peers))
for i, peer := range gc.peers {
cc := pb.ConfChange{Type: pb.ConfChangeAddNode, NodeID: peer.ID, Context: peer.Context}
data, err := cc.Marshal()
if err != nil {
panic("unexpected marshal error")
}
ents[i] = pb.Entry{Type: pb.EntryConfChange, Term: 1, Index: uint64(i + 1), Data: data}
}
r.raftLog.append(ents...)
r.raftLog.committed = uint64(len(ents))
for _, peer := range gc.peers {
r.addNode(peer.ID)
}
}
// Set the initial hard and soft states after performing all initialization.
group.prevSoftSt = r.softState()
group.prevHardSt = r.HardState
close(gc.done)
case gr := <-mn.rmgroupc:
delete(groups, gr.id)
delete(rds, gr.id)
close(gr.done)
case mm := <-mn.propc:
// TODO(bdarnell): single-node impl doesn't read from propc unless the group
// has a leader; we can't do that since we have one propc for many groups.
// We'll have to buffer somewhere on a group-by-group basis, or just let
// raft.Step drop any such proposals on the floor.
mm.msg.From = mn.id
group = groups[mm.group]
group.raft.Step(mm.msg)
case mm := <-mn.recvc:
group = groups[mm.group]
if _, ok := group.raft.prs[mm.msg.From]; ok || !IsResponseMsg(mm.msg) {
group.raft.Step(mm.msg)
}
case mcc := <-mn.confc:
group = groups[mcc.group]
if mcc.msg.NodeID == None {
group.raft.resetPendingConf()
select {
case mcc.ch <- pb.ConfState{Nodes: group.raft.nodes()}:
case <-mn.done:
}
break
}
switch mcc.msg.Type {
case pb.ConfChangeAddNode:
group.raft.addNode(mcc.msg.NodeID)
case pb.ConfChangeRemoveNode:
group.raft.removeNode(mcc.msg.NodeID)
case pb.ConfChangeUpdateNode:
group.raft.resetPendingConf()
default:
panic("unexpected conf type")
}
select {
case mcc.ch <- pb.ConfState{Nodes: group.raft.nodes()}:
case <-mn.done:
}
case <-mn.tickc:
// TODO(bdarnell): instead of calling every group on every tick,
// we should have a priority queue of groups based on their next
// time-based event.
for _, g := range groups {
g.raft.tick()
rd := g.newReady()
if rd.containsUpdates() {
rds[g.id] = rd
}
}
case readyc <- rds:
// Clear outgoing messages as soon as we've passed them to the application.
for g := range rds {
groups[g].raft.msgs = nil
}
rds = map[uint64]Ready{}
advancec = mn.advancec
case advs := <-advancec:
for groupID, rd := range advs {
group, ok := groups[groupID]
if !ok {
continue
}
group.commitReady(rd)
// We've been accumulating new entries in rds which may now be obsolete.
// Drop the old Ready object and create a new one if needed.
delete(rds, groupID)
newRd := group.newReady()
if newRd.containsUpdates() {
rds[groupID] = newRd
}
}
advancec = nil
case ms := <-mn.status:
ms.ch <- getStatus(groups[ms.group].raft)
case <-mn.stop:
close(mn.done)
return
}
if group != nil {
rd := group.newReady()
if rd.containsUpdates() {
rds[group.id] = rd
}
}
}
}